Practical Reinforcement Learning For MPC: Learning from sparse objectives in under an hour on a real robot